[LLVM][CodeGen][SVE] List MVTs that are desirable for extending loads. (#149153)

Extend AArch64TargetLowering::isVectorLoadExtDesirable to specify the
set of MVT for which load extension is desirable.

Fixes #148939

Author: paulwalker-arm

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -7641,7 +7641,7 @@ SDValue DAGCombiner::visitAND(SDNode *N) {
 
     if (SDValue(GN0, 0).hasOneUse() &&
         isConstantSplatVectorMaskForType(N1.getNode(), ScalarVT) &&
-        TLI.isVectorLoadExtDesirable(SDValue(SDValue(GN0, 0)))) {
+        TLI.isVectorLoadExtDesirable(SDValue(N, 0))) {
       SDValue Ops[] = {GN0->getChain(),   GN0->getPassThru(), GN0->getMask(),
                        GN0->getBasePtr(), GN0->getIndex(),    GN0->getScale()};
 
@@ -15745,7 +15745,7 @@ SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
   // fold (sext_inreg (masked_gather x)) -> (sext_masked_gather x)
   if (auto *GN0 = dyn_cast<MaskedGatherSDNode>(N0)) {
     if (SDValue(GN0, 0).hasOneUse() && ExtVT == GN0->getMemoryVT() &&
-        TLI.isVectorLoadExtDesirable(SDValue(SDValue(GN0, 0)))) {
+        TLI.isVectorLoadExtDesirable(SDValue(N, 0))) {
       SDValue Ops[] = {GN0->getChain(),   GN0->getPassThru(), GN0->getMask(),
                        GN0->getBasePtr(), GN0->getIndex(),    GN0->getScale()};
 

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -6439,7 +6439,9 @@ bool AArch64TargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
     }
   }
 
-  return true;
+  EVT PreExtScalarVT = ExtVal->getOperand(0).getValueType().getScalarType();
+  return PreExtScalarVT == MVT::i8 || PreExtScalarVT == MVT::i16 ||
+         PreExtScalarVT == MVT::i32 || PreExtScalarVT == MVT::i64;
 }
 
 unsigned getGatherVecOpcode(bool IsScaled, bool IsSigned, bool NeedsExtend) {

llvm/test/CodeGen/AArch64/sve-intrinsics-ldst-ext.ll

@@ -231,3 +231,27 @@ define <vscale x 8 x i64> @sload_8i8_8i64(ptr %a) {
   %aext = sext <vscale x 8 x i8> %aval to <vscale x 8 x i64>
   ret <vscale x 8 x i64> %aext
 }
+
+; Ensure we don't try to promote a predicate load to a sign-extended load.
+define <vscale x 16 x i8> @sload_16i1_16i8(ptr %addr) {
+; CHECK-LABEL: sload_16i1_16i8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ldr p0, [x0]
+; CHECK-NEXT:    mov z0.b, p0/z, #-1 // =0xffffffffffffffff
+; CHECK-NEXT:    ret
+  %load = load <vscale x 16 x i1>, ptr %addr
+  %zext = sext <vscale x 16 x i1> %load to <vscale x 16 x i8>
+  ret <vscale x 16 x i8> %zext
+}
+
+; Ensure we don't try to promote a predicate load to a zero-extended load.
+define <vscale x 16 x i8> @zload_16i1_16i8(ptr %addr) {
+; CHECK-LABEL: zload_16i1_16i8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ldr p0, [x0]
+; CHECK-NEXT:    mov z0.b, p0/z, #1 // =0x1
+; CHECK-NEXT:    ret
+  %load = load <vscale x 16 x i1>, ptr %addr
+  %zext = zext <vscale x 16 x i1> %load to <vscale x 16 x i8>
+  ret <vscale x 16 x i8> %zext
+}